Page 111 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

Table 3 – List of activities and acronyms
(5)
h t = tanh(C t ).o t
Activity name Acronym
Turning&Milling-Machine TMM
5. RESULTS Turning&MillingQ.C. TMQC
LaserMarking-Machine LMM
Keras [23] was used for the implementation, which is a Python RoundGrinding-Machine RGM
library that allows building models of deep learning networks. RoundQ.C. RQC
The implementation parameters of the LSTM network are FinalInspectionQ.C. FIQC
presented in Table 2. Packing PACK
TurningQ.C. TQC
Table 2 – Configuration parameters of the LSTM neural GrindingRework-Machine GRM
network GrindingRework GR
WireCut-Machine WCM
Parameter Value Fix-Machine FM
epochs 500 NitrationQ.C. NQC
batch size 20
optimizer Adam Table 4 – An extract of the prediction from LSTM
loss categorical_crossentropy
No. Input Target Activity Output Output
LSTM units 50 Activity Activity 1 Activity 2
1 TMQC LMM7 | LPM1 | TMM4 LMM7 LPM1
The LSTM neural network was trained with an event log 2 PACK FIQC | FM15 FIQC
described in Section 3. This event log includes 255 traces of 3 GRM27 FIQC FIQC
4 GR LPM1 | TMQC LPM1
the business process model. There are 56 different activities 5 WCM18 TQC TQC
contained in the log. The number of sequences identified 6 RGM19 RGM12 | FIQC RGM12
7 NQC TMM5 | TMQC TMM5 TMQC
during the network training was 4541. The LSTM network 8 RQC PACK | FIQC PACK FIQC
accepts as input data an activity, in order to predict the 9 FM15 PACK PACK TMQC
10 FGM26 FIQC PACK MM14
next activity of the sequence. The neural network was
configured to predict three outputs per instance, ordered one that was not predicted. For instance, in the first case, were
by a higher to lower probability. The objective is to know predicted the LMM7 and LPM1 activities, but not the TMM4.
the prediction capacity of the neural network of the next However, in these cases, the next activity that is predicted is
activity. The algorithms and datasets can be accessed at the one with the highest probability. Furthermore, in the case
\http://dx.doi.org/10.17632/trskzyg3j9.1. number 9 in which the prediction obtain the desired activity
Table 3 summarizes the activities in the event log and but one of them was not expected in the target. In this instance,
their acronyms. The name of the activities in the table using the FM15 as input, it was expected that the LSTM throw
are acronyms from the real name included in the event as output only the PACK, but the TMQC was also included as
log. For instance: "Turning & MillingQ.C." (TMQC), a response. At last, in the case number 10, the target activity
"LaserMarking-Machine7" (LMM7). is FIQC, but the LSTM network predicts two activities that
Table 4 presents an extract of the results obtained in the do not match with activity what was expected.
prediction of the neural network using the Event Log
presented before. In the column "Input Activity" it is 6. RELATED WORK
mentioned the activity used as a new input for the LSTM
network in the prediction process. The "Target Activity" The development of technological solutions for event log
is the expected activity (or activities) for the corresponding analysis for business process discovery using the principles of
input activity, that is, the activities with the highest probability data mining has been previously studied in [6, 12]. The most
of prediction by the neural network, based on the weights of relevant proposals that are related to the approach proposed
each activity. Each row in the table shows a case of prediction in this research work are discussed in this section. However,
of the next activity from the input one. The "Output Activity" existing techniques are not able to predict at runtime the next
column presents the activities that the LSTM neural network activities that are going to be executed in a business process.
predicted from the input activity. We expect that techniques based on LSTM neural networks,
like the proposed in this work, can also be of help in the
The test carried out on the trained LSTM network shows that discovery of business process models.
it has the capacity to predict the next activity of a business There are a few approaches using patterns and statistical
process model. For the cases number 3, 5, 7 and 8, the models to predict activities in business processes. The
network was able to predict the exact next activity. For approach described in [24], aims at identifying partial
instance, in the third case, receiving the GRM27 as input, business process models to be used for training predictive
the LSTM network was able to predict the expected FIQC models. It infers two types of predictive models. The
(the output activity is included in the target activity list, with first model is used to identify frequent partial processes
the highest probability). In other cases, as the number 1, 2, 4 in form of frequent activity sequences, the sequences are
and 6, the most of the target activities were identified, missing extracted using a frequent pattern mining algorithm and are

– 95 –